Voice Conversion For Song Based On Latent Variable Model

Voice conversion is a important branch of speech signal processing. In recent years, domestic and foreign scholars have achieved many research results in the field of voice conversion. This paper presents a voice conversion that makes the voice of the source singer similar to the target singer, while keeping the lyrics unchanged. The technique of voice conversion is expected to have a great prospect in the film dubbing, physiology, entertainment and simultaneous interpretation.The traditional voice conversion is commonly how to establish a mapping from the voice personality characteristics of the source speaker and the target speaker. The basic idea of this method is that extract the personality characteristics of the source speaker and the target speaker voices in a unified voice- synthesis model, and find the mapping of those personality characteristics. This paper research the extraction of the voice feature, we design two voice conversion method:(1)We will be decomposed into songs to content and the personality characteristics by building a latent variable models. Last the source personality characteristics will replace to the target;(2)We presents a voice conversion technique using Deep belief networks(DBN) to build high-order eigen spaces of the source/target speakers, where it is easier to convert the source speech to the target speech than in the traditional cepstrum space.Finally, this paper use two algorithms to conducted experiments and results analysis, Experiments show the song conversion is a feasible.